Function GenerateDeconv_Simt

dnn/scripts/cutlass_generator/generator.py:712–768

(args)

def GenerateDeconv_Simt(args):
    operations = []

    layouts = [(LayoutType.TensorNC4HW4, LayoutType.TensorK4RSC4)]

    math_instructions = [
        MathInstruction(
            [1, 1, 4],
            DataType.s8,
            DataType.s8,
            DataType.s32,
            OpcodeClass.Simt,
            MathOperation.multiply_add,
        )
    ]

    dst_layouts = [LayoutType.TensorNC4HW4]

    dst_types = [DataType.s8]

    use_special_optimization = SpecialOptimizeDesc.DeconvDoubleUpsampling

    min_cc = 61
    max_cc = 1024

    for math_inst in math_instructions:
        for layout in layouts:
            for dst_type, dst_layout in zip(dst_types, dst_layouts):
                tile_descriptions = [
                    TileDescription(
                        [32, 128, 32], 2, [1, 2, 1], math_inst, min_cc, max_cc
                    ),
                    TileDescription(
                        [16, 128, 16], 2, [1, 2, 1], math_inst, min_cc, max_cc
                    ),
                    TileDescription(
                        [16, 128, 16], 1, [1, 1, 1], math_inst, min_cc, max_cc
                    ),
                    TileDescription(
                        [16, 64, 8], 2, [1, 1, 1], math_inst, min_cc, max_cc
                    ),
                ]
                operations += GenerateConv2d(
                    ConvType.Convolution,
                    ConvKind.Dgrad,
                    tile_descriptions,
                    layout[0],
                    layout[1],
                    dst_layout,
                    dst_type,
                    min_cc,
                    32,
                    32,
                    32,
                    use_special_optimization,
                )
    return operations